High precision classification of hot rolled strip steel surface defects using dual path features and entropy attention fusion

In the context of industrial hot-rolled strip steel surface defect detection, where the demands for real-time performance and classification accuracy are paramount, we present EAF-DenseNet121-a lightweight, enhanced model that incorporates edge-entropy attention mechanisms. At the inception of the DenseNet121 architecture, we incorporate a learnable Sobel-based edge extraction branch, which is designed to adaptively delineate defect contours with precision. We have designed an Entropy-Attention Fusion (EAF) module to further refine the model's performance. This module constructs a four-dimensional tensor, integrating the primary feature map, edge map, and their corresponding local entropy maps. By applying dual-path channel-wise and spatial attention, we achieve a weighted fusion of information, thereby enriching the feature representation. The EAF module replaces three pivotal convolutional layers within the DenseNet framework-immediately following the initial convolution and subsequent to the first and second Transition layers. This replacement enhances feature representation with a negligible increase in additional parameters, leading to a substantial improvement in defect recognition and classification accuracy. Our experimental results, obtained on the NEU-DET dataset, reveal that the enhanced model achieves a classification accuracy of 99.17%, representing an improvement of 2.78% over the baseline. Furthermore, on the GC10-DET dataset, the model achieves a classification accuracy of 82.89%, further validating its strong generalization capabilities.